For PHS 7045
The University of Utah
2024-12-12
Hill-of-vision: a graphical representation of retinal sensitivity, “peaks” represent areas of high sensitivity.
We would like to use this hill-of-vision map to calculate standard sensitivity loss
Conservative method: spatial interpolation
Spatial interpolation estimates sensitivity for a single virtual patient’s testing results using Kriging.
However, this approach may lose some population-level information and is harder to interpret.
An interpretable age-adjusted prediction model would be much better!
Works with multiple exam types (e.g., Mesopic, Cyan, Red, and Cyan-Red difference).
Provides predicted sensitivity values and visualizations with parallel computing.
Kriging uses the best linear unbiased prediction based on both the spatial structure and the measurements, so it’s not guaranteed to be the best prediction, especially in areas with sparse test points.
Objective: Predict retinal sensitivity based on patient-specific characteristics like age and test location.
Models Used: Linear Mixed Model (LMM), Bayesian Quantile Regression (BQR) and Random Forest (RF).
Performance comparison: mean absolute error (MAE) and mean absolute calibration error (MACE).
SensForFit()PredictCompare()
PredictNormal_lmm(), PredictNormal_bqr(), PredictNormal_rf()PredictSensitivity()VisualRetina()data("refMes")
## data preparation
refMes <- SensForFit(dt = refMes, examcol = "Examtype", idcol = "Patient", agecol = "Age", senscol = "MeanSens", k = 10)
## model comparison
(res.tab <- PredictCompare(dt = refMes, exam = "Mesopic", CalibSplit = 0.2, coverage = 0.95))
## prediction for a future patient aged 77 (e.g. bqr is the best)
pred_BQR_77 <- PredictSensitivity(model = "BQR", age = 77, dt = refMes)
## hill-of-vision for new patient
VisualRetina(pred_sens = pred_BQR_77, exam = "Mesopic")| Model | MAE | MACE | |
|---|---|---|---|
| lmm_results | LMM | 1.43 | 0.03 |
| bqr_results | BQR | 1.70 | 0.02 |
| rf_results | RF | 1.53 | 0.03 |
Other Models vs. Interpolation
Models provide insights into age and test-location effects on sensitivity.
No need to predict complete hill-of-vision.
Models can account for variability across patients and test sites.
Further exploration is needed to develop a robust conformal prediction framework, especially for machine learning methods.
Yingjia Wei